Solving Locomotor Problems

Problem solving is integral to functional locomotion. To successfully navigate the environment, locomotion must be continually modified to suit changes in local conditions—different surfaces (e.g., walking on pavement or sand), changes in layout (walking uphill or over flat ground), and obstacles in the path (clutter, elevations, other moving agents). No step is ever repeated in exactly the same way or under exactly the same conditions. Thus, everyday locomotion requires agents to solve the problem of adapting to moment-to-moment changes in body-environment relations.

 

In one line of research, we use robots as a model system to show that path variability and low penalty for error, prime characteristics of infant walking, are crucial features in learning to solve locomotor problems. For example, we showed that training simulated robots (evolutionary learning algorithms) on real infant walking paths led to more wins in robot soccer (“Robocup”) than training on traditional training paths (geometric shapes). Moreover, robots trained on more variable infant paths with minimal penalty for falling won more soccer matches than robots trained on less variable paths or with a high penalty for falling. Our findings indicate that infants provide a natural training set for learning functional locomotion. The optimal training paths were best captured by a specific random distribution called a “Levy walk” that ensures optimal foraging behavior.


In a second line of research, we use AI to study how infants solve the problem of incorporating external objects into a plan of action and coordinate their bodies with a goal and a tool. Specifically, we used deep machine learning to discover how infants use a handrail as a tool to augment balance and the strategies they learned to successfully coordinate their arms and legs to “cruise” over gaps in a handrail.

 

My integrative analytic approach provides a novel, precise, and objective tool for studying the development of locomotor problem solving .